README.md

---
license: cc-by-sa-4.0
task_categories:
- text-retrieval
- sentence-similarity
language:
- en
tags:
- wikipedia
- embeddings
- splade
- bge-m3
- dense-retrieval
- sparse-retrieval
- hybrid-search
size_categories:
- 1M<n<10M
---

# Wikipedia English with SPLADE and BGE-M3

Pre-computed SPLADE sparse and BGE-M3 dense embeddings for 6.4M English Wikipedia articles.

## Dataset Description

- **Source**: [HuggingFaceFW/clean-wikipedia](https://huggingface.co/datasets/HuggingFaceFW/clean-wikipedia) (English)
- **Size**: 6,406,711 documents
- **Format**: 7 Parquet files (~1M records each)
- **Total Size**: ~15-20GB
- **Embeddings**: 
  - **Sparse**: [SPLADE PP](https://huggingface.co/prithivida/Splade_PP_en_v1) - ~265 non-zero dims
  - **Dense**: [BGE-M3](https://huggingface.co/BAAI/bge-m3) - 1024 dimensions

## Direct Usage

HuggingFace automatically discovers parquet files. You can load this dataset directly:

```python
from datasets import load_dataset

# Stream the entire dataset (recommended for large dataset)
dataset = load_dataset("Sicheng-Chroma/wikipedia-en-splade-bge", streaming=True)

# Load specific split/chunk
dataset = load_dataset("Sicheng-Chroma/wikipedia-en-splade-bge", 
                       data_files="train-00000-of-00007.parquet")

# Load multiple chunks
dataset = load_dataset("Sicheng-Chroma/wikipedia-en-splade-bge", 
                       data_files=["train-00000-of-00007.parquet", "train-00001-of-00007.parquet"])
```

## Dataset Structure

Each row contains:

| Field | Type | Description |
|-------|------|-------------|
| `text` | string | Full Wikipedia article text |
| `title` | string | Article title |
| `url` | string | Wikipedia URL |
| `sparse_embedding_indices` | list[int32] | SPLADE indices (non-zero positions) |
| `sparse_embedding_values` | list[float32] | SPLADE values (weights) |
| `dense_embedding` | list[float32] | BGE-M3 1024-dim dense vector |

## Example Usage

```python
from datasets import load_dataset

# Load the dataset (streaming recommended for large dataset)
ds = load_dataset("Sicheng-Chroma/wikipedia-en-splade-bge", streaming=True)

# Iterate through articles
for article in ds['train']:
    print(f"Title: {article['title']}")
    print(f"URL: {article['url']}")
    print(f"Text preview: {article['text'][:200]}...")
    print(f"Sparse embedding: {len(article['sparse_embedding_indices'])} non-zero dims")
    print(f"Dense embedding: {len(article['dense_embedding'])} dims")
    break

# Load into memory (for smaller chunks)
dataset = load_dataset("Sicheng-Chroma/wikipedia-en-splade-bge", 
                       data_files="train-00000-of-00007.parquet")
print(f"Loaded {len(dataset['train'])} articles")
```

## Files

The dataset consists of 7 parquet files in standard HuggingFace format:
- `train-00000-of-00007.parquet` - First 1M articles
- `train-00001-of-00007.parquet` - Next 1M articles
- ...
- `train-00006-of-00007.parquet` - Final ~400K articles

## License

CC BY-SA 4.0 (inherited from Wikipedia)